TensorFlow 1.12 and Horovod support (#138)

Author: mvsusp

benchmarks/README.md

@@ -3,33 +3,64 @@
 This folder contains the TF training scripts https://github.com/tensorflow/benchmarks/tree/master/scripts/tf_cnn_benchmarks.
 
 ## Basic usage
-**execute_tensorflow_training.py** uses SageMaker python sdk to start a training job. It takes the following parameters:
+**execute_tensorflow_training.py train** uses SageMaker python sdk to start a training job. 
 
-- role: SageMaker role used for training
-- region: SageMaker region
-- py-versions: py2 or py3 or "py2, py3"
-- instance-types: A list of SageMaker instance types, for example 'ml.p2.xlarge, ml.c4.xlarge'. Use 'local' for local mode training.
-- checkpoint-path: The S3 location where the model checkpoints and tensorboard events are saved after training 
+```bash
+./execute_tensorflow_training.py train --help
+Usage: execute_tensorflow_training.py train [OPTIONS] [SCRIPT_ARGS]...
+
+Options:
+  --framework-version [1.11.0|1.12.0]
+                                  [required]
+  --device [cpu|gpu]              [required]
+  --py-versions TEXT
+  --training-input-mode [File|Pipe]
+  --networking-isolation / --no-networking-isolation
+  --wait / --no-wait
+  --security-groups TEXT
+  --subnets TEXT
+  --role TEXT
+  --instance-counts INTEGER
+  --batch-sizes INTEGER
+  --instance-types TEXT
+  --help                          Show this message and exit.
 
-Any unknown arguments will be passed to the training script as additional arguments.
+```
+**execute_tensorflow_training.py generate_reports** generate benchmark reports. 
 
 ## Examples:
 
 ```bash
-./execute_tensorflow_training.py -t local -r SageMakerRole --instance-type local  --num_epochs 1 --wait
-
-./execute_tensorflow_training.py -t local -r SageMakerRole --instance-type ml.c4.xlarge, ml.c5.xlarge  --model resnet50
+#!/usr/bin/env bash
 
+./execute_tensorflow_training.py train \
+--framework-version  1.11.0 \
+--device gpu \
+\
+--instance-types ml.p3.2xlarge \
+--instance-types ml.p3.8xlarge \
+--instance-types ml.p3.16xlarge \
+--instance-types ml.p2.xlarge \
+--instance-types ml.p2.8xlarge \
+--instance-types ml.p2.16xlarge \
+\
+--instance-counts 1 \
+\
+--py-versions py3 \
+--py-versions py2 \
+\
+--subnets subnet-125fb674  \
+\
+--security-groups sg-ce5dd1b4  \
+\
+--batch-sizes 32 \
+--batch-sizes 64 \
+--batch-sizes 128 \
+--batch-sizes 256 \
+--batch-sizes 512 \
+\
+-- --model resnet32 --num_epochs 10 --data_format NHWC --summary_verbosity 1 --save_summaries_steps 10 --data_name cifar10
 ```
 
 ## Using other models, datasets and benchmarks configurations
 ```python tf_cnn_benchmarks/tf_cnn_benchmarks.py --help``` shows all the options that the script has.
-
-
-## Tensorboard events and checkpoints
-
-Tensorboard events are being saved to the S3 location defined by the hyperparameter checkpoint_path during training. That location can be overwritten by setting the script argument ```checkpoint-path```:
-
-```bash
-python execute_tensorflow_training.py ... --checkpoint-path s3://my/bucket/output/data
-```

benchmarks/benchmarks

@@ -0,0 +1 @@
+Subproject commit ec056be57f189ec96611a58e8dc5562a6d620139